Artificial Intelligence Off-Policy Evaluation of Community Clinic Fever Triage Under Unmeasured Antimalarial Co-Administration: A Healthcare Doubly Robust Regret Evaluation

Community clinic fever triage in malaria-endemic settings is a sequential decision problem in which the same patient may receive an antimalarial and a triage recommendation, and the antimalarial is recorded inconsistently across registers. Off-policy evaluation of triage policies from observational registers therefore faces unmeasured co-administration: the treatment that the policy would recommend and the treatment that was actually given are entangled. The estimator combines an inverse-propensity-weighted value with a regression-based value and is shown to be consistent when either the propensity model or the outcome model is correctly specified, but not necessarily both. A formal analysis establishes that unmeasured co-administration enters the regret as a bias term proportional to the conditional covariance between the co-administration indicator and the triage action, scaled by the inverse of the propensity score. The bias is bounded under a sensitivity parameter that quantifies the departure from sequential ignorability. The estimator is presented as a methodological proposal for register-based triage evaluation; no empirical application is reported. The analysis implies that regret comparisons between triage policies are informative only when the sensitivity parameter is reported alongside the point estimate, and that registers which record antimalarial dispensing separately from triage recommendation permit tighter bounds than registers which do not.